⇄QKD networks: architectures, routing and deployments

Network architectures

A single QKD link spans at most a few hundred kilometres. Networks extend reach and connect many users:

In a trusted-relay network, to send key $K$ from A to C via B, A and B share $K_{AB}$, B and C share $K_{BC}$; B receives $K\oplus K_{AB}$ from A, decrypts, re-encrypts with $K_{BC}$ and forwards. Each hop therefore consumes as much link key as is delivered.

Physical-layer key-rate model

For a decoy-state BB84 link of length $L$ with repetition rate $f_{\rm rep}$ the secret-key rate per second is approximately

so the rate falls exponentially with distance (dB loss is linear in $L$) until detector dark counts dominate. The optimal mean photon number $\mu$ depends on distance and noise, and can be tuned dynamically. Co-existing with classical channels in the same fibre adds Raman noise that raises QBER; wavelength-division multiplexing (WDM) planning must keep the quantum channel in a quiet band and filter tightly.

Key management layer

Applications do not consume raw QKD output. A key management system (KMS) stores keys in pools (per link and per pair of endpoints), authenticates them, and delivers them through standard interfaces. ETSI GS QKD 004 and 014 define application-facing key-delivery APIs; ITU-T Y.3800-series recommends QKD-network architectures. Because key generation is slow and bursty while demand varies, buffering and key-aging policies matter.

Resource allocation, routing and relay placement

• Resource allocation. Share limited key rate, wavelengths (WDM) and detectors among competing users; formulated as optimisation with key-pool and latency constraints.
• Routing in trusted-relay networks. Metrics: residual key in pools, link key rate, hop count, latency and security level. Multi-path routing splits a key over disjoint paths and XORs the parts, so that one compromised relay does not reveal the key (key aggregation).
• Entanglement-based routing. In repeater networks routing selects paths for entanglement swapping with fidelity and memory-time constraints.
• Trusted-relay-node (TRN) placement. Choose a minimum number of relay sites so that every pair of users is connected with links no longer than a maximum QKD range, subject to cost, terrain, security and redundancy. This is a coverage/connectivity problem (NP-hard), solved by integer programming, greedy and heuristic methods.

Network-level attacks

Beyond link attacks (see Security): relay compromise (keys exposed at trusted nodes), denial of service (fibre cuts, jamming, noise injection that pushes QBER over the abort threshold), key-pool depletion and management-plane attacks. Defences: physical protection, redundancy, multi-path key delivery, intrusion monitoring, rate limiting and combining QKD keys with post-quantum authentication.

Machine learning for QKD networks (research)

Active research areas: reinforcement learning for key-pool and routing decisions; supervised learning for QBER and channel prediction and adaptive parameter selection (e.g. $\mu$, decoy intensities); unsupervised anomaly detection for attacks and drift; federated/transfer learning across sites; graph neural networks for topology and placement. Security caution: ML components expand the attack surface and must not affect the security-critical parts of the protocol.

Selected deployments

Dates and scales below are indicative; consult primary sources.